feat(studio-cp,oabctl): build a default manifest when none exists yet — the actual fix for studio#111 - #113
Closed
brettchien wants to merge 1 commit into
Conversation
… — the actual fix for #111 provision_from_library (backing deploy_provision / the console's "+ New fleet" flow) now checks load_manifest first: if a manifest is already stored, behavior is unchanged (redeploy patches image/bundle_from and re-applies). If none exists, it builds a fresh OABServiceManifest (build_default_manifest) using #112's zero-prompt defaults (VPC/subnet/SG via default_networking, resources 256/512, FARGATE/X86_64, empty secrets — see #111's comment thread for why empty secrets is valid and not a gap this needs to solve) and applies it via provision_manifest, a new oabctl:: studio_api helper that mirrors provision() but takes a structured manifest instead of YAML text (keeps serde_yaml an oabctl-internal detail). This is the actual fix: the console's "+ New fleet" wizard can now create a genuinely new agent end-to-end, not just redeploy an agent someone already created via the CLI. Once merged, k8s's deploy_provision dispatch (#104) lands on this same branch point — load_manifest is provider-agnostic (S3 key, not ECS-specific), so the create-vs-redeploy check doesn't need to change for k8s; only the "build a fresh manifest" + "apply it" halves need a Runtime::Kubernetes(...) branch alongside this Runtime::Ecs(...) one. configFrom for the new manifest points at artifacts/{ns}/{name}/config.toml — the same key Bundle::artifact_objects already uploads a copy of the composed config.toml to, and the same convention oabctl create's wizard uses. Unit-tested (default_config_from_uri_matches_artifact_objects_key). Manually verified every field against crates/oabctl/src/manifest.rs's struct definitions (OABServiceManifest/Metadata/Spec/Resources/Runtime/ EcsRuntime/EcsNetworking) and OABServiceManifest::validate()'s requirements (apiVersion "oab.dev/v2", kind "OABService", CPU "256" is in VALID_ECS_CPU, capacityProvider "FARGATE" is valid) — this is the riskiest change this session (real infra creation), so more care than usual went into checking it by hand given the environment's known aws-sdk-ec2 OOM constraint prevented a local cargo check. CI is the gate. Ref: studio#111.
This was referenced Aug 27, 2026
brettchien
added a commit
that referenced
this pull request
Aug 27, 2026
…104) t_provision now branches on a new provider arg (default "aws", unchanged behavior): "k8s" calls provision_from_library_k8s (#114) with context/ expected_principal taken as direct args, not resolved via a fleet name — the console's identity form already collects context/namespace/service- account directly (#108/#109), and for a brand-new fleet there's no existing K8sFleetBinding to look up anyway. Fleet-scoped k8s lookup (for redeploying into an already-known k8s fleet) is left as explicit future work, not needed for this dispatch to exist. Also updates deploy_provision's tool description, which was stale after #113 (redeploy no longer requires an existing manifest) and needed the new provider/context/expected_principal params documented. NOTE — branch lineage: this stack (#112→#113→#114→this) was cut from `main` before #104's original stack (#105-110) merged, not from #110 — so K8sFleetBinding.expected_principal (originally #107) is re-added here too. Identical field in both places; trivial merge conflict to resolve whenever both land, flagging explicitly rather than silently duplicating without a note. With this, deploy_provision fully supports k8s end-to-end (provisioning side) — the remaining piece for full onboarding is unblocking console's k8s "Next" button (#108/#109's placeholder), a separate follow-up. Ref: studio#104.
Contributor
Author
brettchien
added a commit
that referenced
this pull request
Aug 28, 2026
…est-creation fix, k8s deploy_provision dispatch) (#116) * feat(oabctl): expose create.rs's AWS-placement defaults for reuse (studio#111) First sub-item of #111 ("+ New fleet" first-creation gap). oabctl create's interactive wizard collects VPC/subnet/security-group among other things before it can build a manifest and apply it — but the VPC/subnet/SG parts are already effectively "sensible defaults", not real interactive choices: select_subnets already auto-picks (private+NAT > private > public, up to 3 AZs) with zero prompting, and "Create new (oab-{name})" is the SG wizard's own default suggestion. Makes `list_vpcs`/`VpcInfo`, `select_subnets`/`SubnetInfo`, `list_security_groups`/`SgInfo` pub (was module-private, module itself was `mod create` — now `pub mod create`), and adds one new function, `default_security_group`, extracting the "create new (oab-{name}), reuse if it already exists" logic `run()` has inline into something callable without the interactive prompt path. No behavior change to `run()`/`oabctl create` — pure visibility change plus one new function built from existing, already-working logic (the AWS calls in `default_security_group` are the same calls `run()`'s inline SG-create branch already makes). This is groundwork only — nothing calls these yet. Next: the console-facing provision path needs a VPC choice too (unlike subnet/SG, there's no zero-prompt default for "which VPC" today — `run()` always asks). Then wiring these into a "build a default manifest, apply via deploy_apply's path" function is the actual fix for #111. Local build/test hit the known aws-sdk-ec2 OOM constraint on this machine (same pre-existing environment issue documented in prior PRs' descriptions) — change is a mechanical visibility change + one function built from already-proven inline logic, hand-verified against the diff. CI is the gate. Ref: studio#111. * feat(oabctl): add default_vpc — zero-prompt VPC pick when the account has exactly one default (studio#111) Follow-up on this same PR: VpcInfo now carries is_default (was folded into the human-readable label only, not usable programmatically). default_vpc() picks the account/region's default VPC when there's exactly one — the closest zero-prompt equivalent to what select_subnets/default_security_group already give for subnet/SG, since run()'s wizard never had a non-interactive default for VPC choice at all. Deliberately errors (not a heuristic guess) when there's zero or more than one default VPC — a caller with an explicit VPC choice (e.g. future per-fleet config, mirroring how fleets.toml already carries region/profile) should skip this and pass that VPC straight to select_subnets/default_security_group instead. Still groundwork — nothing calls this yet. Ref: studio#111. * feat(oabctl): add default_networking — SdkConfig-only entry point for VPC/subnet/SG defaults (studio#111) Follow-up on this same PR. Wraps default_vpc + select_subnets + default_security_group behind one function that takes an aws_config:: SdkConfig, not an Ec2Client — so studio-cp (which doesn't depend on aws-sdk-ec2 directly) can reach it without adding that dependency, keeping Ec2Client an oabctl-internal detail (same "RuntimeDriver is the only layer with vendor terms" boundary ADR-2 already established elsewhere). Still groundwork — nothing calls this yet. Next: build_default_manifest in studio-cp, calling this + spec defaults (resources 256/512, empty secrets, FARGATE/X86_64), then wire provision_from_library to branch create-vs- redeploy based on whether load_manifest finds a stored manifest. Ref: studio#111. * feat(studio-cp,oabctl): build a default manifest when none exists yet — the actual fix for #111 provision_from_library (backing deploy_provision / the console's "+ New fleet" flow) now checks load_manifest first: if a manifest is already stored, behavior is unchanged (redeploy patches image/bundle_from and re-applies). If none exists, it builds a fresh OABServiceManifest (build_default_manifest) using #112's zero-prompt defaults (VPC/subnet/SG via default_networking, resources 256/512, FARGATE/X86_64, empty secrets — see #111's comment thread for why empty secrets is valid and not a gap this needs to solve) and applies it via provision_manifest, a new oabctl:: studio_api helper that mirrors provision() but takes a structured manifest instead of YAML text (keeps serde_yaml an oabctl-internal detail). This is the actual fix: the console's "+ New fleet" wizard can now create a genuinely new agent end-to-end, not just redeploy an agent someone already created via the CLI. Once merged, k8s's deploy_provision dispatch (#104) lands on this same branch point — load_manifest is provider-agnostic (S3 key, not ECS-specific), so the create-vs-redeploy check doesn't need to change for k8s; only the "build a fresh manifest" + "apply it" halves need a Runtime::Kubernetes(...) branch alongside this Runtime::Ecs(...) one. configFrom for the new manifest points at artifacts/{ns}/{name}/config.toml — the same key Bundle::artifact_objects already uploads a copy of the composed config.toml to, and the same convention oabctl create's wizard uses. Unit-tested (default_config_from_uri_matches_artifact_objects_key). Manually verified every field against crates/oabctl/src/manifest.rs's struct definitions (OABServiceManifest/Metadata/Spec/Resources/Runtime/ EcsRuntime/EcsNetworking) and OABServiceManifest::validate()'s requirements (apiVersion "oab.dev/v2", kind "OABService", CPU "256" is in VALID_ECS_CPU, capacityProvider "FARGATE" is valid) — this is the riskiest change this session (real infra creation), so more care than usual went into checking it by hand given the environment's known aws-sdk-ec2 OOM constraint prevented a local cargo check. CI is the gate. Ref: studio#111. * feat(studio-cp,oabctl): k8s deploy_provision dispatch (studio#104, resumed after #111) Resumes #104's k8s deploy_provision work now that #111 (#112/#113) gave both drivers a shared, provider-agnostic create-vs-redeploy branch point — load_manifest is just an S3 key lookup, it doesn't care which Runtime variant a stored manifest holds. - oabctl::studio_api::provision_k8s: provision_manifest's k8s counterpart, applies through K8sDriver instead of EcsDriver. Takes two separate credential contexts (aws_config for the S3 bundle carrier — hooks.pre_seed is provider-agnostic, still S3 regardless of runtime — and a kubeconfig context for the actual apply) since k8s provisioning genuinely needs both simultaneously, unlike the ECS path where one SdkConfig covers everything. - studio-cp::build_default_k8s_manifest: Runtime::Kubernetes counterpart to build_default_manifest. No VPC/subnet/SG (ECS-only networking concept); service_account comes from K8sFleetBinding.expected_principal when it names one (k8s_service_account_from_principal extracts the bare name from the system:serviceaccount:<ns>:<name> form the New Fleet wizard's service- account picker writes — KubernetesRuntime.service_account wants the bare name, verified against k8s_driver.rs's own build_deployment_wires_service_account_and_node_selector test). - studio-cp::provision_from_library_k8s: provision_from_library's k8s counterpart. Compose/bundle-upload logic is duplicated rather than shared for now (deliberate — avoids reworking provision_from_library's shape again while it's still unmerged; worth revisiting once both paths are proven). Redeploy preserves the stored manifest's k8s runtime config, only bumps the image, same guarantee the AWS path gives. Nothing calls provision_from_library_k8s yet — oab-mcp's deploy_provision tool still needs a provider param to dispatch to it (the OabMcp struct also has no k8s-fleet-binding awareness yet to resolve context/expected_principal from). That wiring is the next piece. Manually verified every field against manifest.rs's struct definitions, same care as #113 given the environment can't locally compile (aws-sdk-ec2 OOM) — CI is the gate. Ref: studio#104, studio#111. * feat(oab-mcp): wire deploy_provision's k8s provider dispatch (studio#104) t_provision now branches on a new provider arg (default "aws", unchanged behavior): "k8s" calls provision_from_library_k8s (#114) with context/ expected_principal taken as direct args, not resolved via a fleet name — the console's identity form already collects context/namespace/service- account directly (#108/#109), and for a brand-new fleet there's no existing K8sFleetBinding to look up anyway. Fleet-scoped k8s lookup (for redeploying into an already-known k8s fleet) is left as explicit future work, not needed for this dispatch to exist. Also updates deploy_provision's tool description, which was stale after provider/context/expected_principal params documented. NOTE — branch lineage: this stack (#112→#113→#114→this) was cut from `main` before #104's original stack (#105-110) merged, not from #110 — so K8sFleetBinding.expected_principal (originally #107) is re-added here too. Identical field in both places; trivial merge conflict to resolve whenever both land, flagging explicitly rather than silently duplicating without a note. With this, deploy_provision fully supports k8s end-to-end (provisioning side) — the remaining piece for full onboarding is unblocking console's k8s "Next" button (#108/#109's placeholder), a separate follow-up. Ref: studio#104.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
The actual fix for #111, stacked on #112 (merge #112 first).
provision_from_library(backingdeploy_provision/ the console's "+ New fleet" flow) now handles both cases:redeploy()patchesimage/bundle_fromand re-applies.OABServiceManifest(build_default_manifest) using feat(oabctl): expose create.rs's AWS-placement defaults for reuse (studio#111) #112's zero-prompt defaults (VPC/subnet/security-group viadefault_networking, resources256/512,FARGATE/X86_64, emptysecrets) and applies it via a newoabctl::studio_api::provision_manifest(mirrorsprovision()but takes a structured manifest instead of YAML text, keepingserde_yamlanoabctl-internal detail).This means the console's "+ New fleet" wizard can create a genuinely new agent end-to-end now — previously it could only redeploy an agent someone had already created via the CLI (
oabctl create), which is the gap #111 traces back to.Why empty
secretsis fine (not a gap this PR needs to solve)Discussed at length in #111's comment thread:
OABServiceManifest.spec.secretsdefaults to empty and that's a valid manifest — it just means no ECS-level Secrets Manager env-var injection happens. Whatever a template's ownconfig.tomlneeds (Discord token, etc.) is the operator's existing responsibility, unchanged from howoabctl createalready works today. Not a new limitation this PR introduces.Where this leaves k8s (#104)
load_manifestis provider-agnostic (an S3 key, not ECS-specific), so the create-vs-redeploy branch point doesn't need to change for k8s — only the two halves (build-fresh-manifest, apply-it) need aRuntime::Kubernetes(...)branch alongside this PR'sRuntime::Ecs(...)one. That's the resumption point for #104'sdeploy_provisionk8s dispatch once this merges.Verification
This is the riskiest change of the whole #104/#111 effort (real AWS infrastructure creation, not just reads/tooling) — the environment's
aws-sdk-ec2OOM constraint (documented in every PR this session) meant no localcargo checkcompleted, so extra care went into manual verification:OABServiceManifestchecked againstcrates/oabctl/src/manifest.rs's actual struct definitions (OABServiceManifest/Metadata/Spec/Resources/Runtime/EcsRuntime/EcsNetworking).OABServiceManifest::validate()'s actual requirements:apiVersionmust be"oab.dev/v2",kindmust be"OABService", CPU"256"is inVALID_ECS_CPU, capacity provider"FARGATE"is in the valid set.default_config_from_uri(the one pure, extractable piece) has a unit test confirming it matches the S3 keyBundle::artifact_objectsalready uploadsconfig.tomlto.CI (
build-test) is the real gate here given the above — flagging explicitly per this session's established practice, not asserting local-verified confidence.Scope note
k8s dispatch itself is out of scope here — tracked in #104, resumes once this merges.
Ref #111. Stacks on #112 — merge #112 first.